Kurdish speech data faces reuse challenges despite fluency
Summary and headline written by AI from the source article. How we work
Researchers evaluated a recently released public dataset of Central Kurdish speech recordings, 35 hours in total, to assess its usability for further development of speech technologies.
The team found several inconsistencies between the dataset’s documentation and its contents. A settings file referenced equipment not actually used in the recordings, and some test data was incorrectly included with the training data. The review also identified a coding error that affects the pronunciation of longer numbers.
Furthermore, the dataset’s online description overstated the achieved results and suggested one voice as generally suitable, a recommendation the researchers caution against given the substantial regional and written differences within the Kurdish language. Central Kurdish is spoken by millions in Iraq, Iran, Syria and Turkey, but lacks the extensive linguistic resources available for languages like English and German, making error detection more difficult.
While the generated voices sound natural, they reflect the reading styles of the three speakers used for data collection and do not fully represent the diversity of everyday Kurdish speech. The researchers note that most of these issues could be resolved with better documentation and data management. However, the dataset's licensing terms, specifically, whether audiobook owners will permit sharing of corrected versions, will determine if future work can build upon this foundation or must begin anew.
